{"product_id":"50-mẹo-claude-code-thực-chiến-từ-thinking-tokens-dến-subagents","title":"50 Mẹo Claude Code Thực Chiến: Từ Thinking Tokens đến Subagents","description":"\n\u003ch2\u003eTại sao cần tips để dùng Claude Code tốt hơn?\u003c\/h2\u003e\n\u003cp\u003eClaude Code không phải chatbox — đây là AI agent với nhiều tính năng nâng cao mà phần lớn developer không khám phá hết. Builder.io đã tổng hợp 50 tips từ cộng đồng developer dùng Claude Code daily, từ những trick đơn giản tiết kiệm vài giây đến những pattern architecture thay đổi cách bạn tiếp cận cả project.\u003c\/p\u003e\n\n\u003cp\u003eBài viết này chắt lọc những mẹo quan trọng nhất, nhóm theo chủ đề để dễ áp dụng ngay.\u003c\/p\u003e\n\n\u003ch2\u003eNhóm 1: Quản lý Thinking Budget\u003c\/h2\u003e\n\n\u003ch3\u003eHierarchy \"think\" keywords\u003c\/h3\u003e\n\u003cp\u003eĐây là một trong những tricks ít người biết nhất của Claude Code. Bạn có thể control mức độ \"suy nghĩ sâu\" của Claude bằng từ khóa:\u003c\/p\u003e\n\n\u003cul\u003e\n  \u003cli\u003e\n\u003ccode\u003ethink\u003c\/code\u003e — Standard thinking, đủ cho tasks thông thường\u003c\/li\u003e\n  \u003cli\u003e\n\u003ccode\u003ethink hard\u003c\/code\u003e — Extended thinking, cho vấn đề có nhiều tradeoffs\u003c\/li\u003e\n  \u003cli\u003e\n\u003ccode\u003ethink harder\u003c\/code\u003e — Deep thinking, cho architecture decisions phức tạp\u003c\/li\u003e\n  \u003cli\u003e\n\u003ccode\u003eultrathink\u003c\/code\u003e — Maximum thinking budget, dành cho legacy code integration, complex algorithms, hay khi cần thoroughness tuyệt đối\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003cp\u003e\u003cstrong\u003eCảnh báo quan trọng:\u003c\/strong\u003e Đừng dùng \u003ccode\u003eultrathink\u003c\/code\u003e cho mọi thứ. Nó tiêu tốn nhiều token và thời gian hơn đáng kể. Match effort với problem — \u003ccode\u003eultrathink\u003c\/code\u003e chỉ khi thực sự cần.\u003c\/p\u003e\n\n\u003ch3\u003eKhi nào dùng gì?\u003c\/h3\u003e\n\u003cul\u003e\n  \u003cli\u003e\"Add a button to this form\" → \u003ccode\u003ethink\u003c\/code\u003e (hoặc không cần keyword)\u003c\/li\u003e\n  \u003cli\u003e\"Refactor this function với better error handling\" → \u003ccode\u003ethink hard\u003c\/code\u003e\n\u003c\/li\u003e\n  \u003cli\u003e\"Design caching strategy cho API này\" → \u003ccode\u003ethink harder\u003c\/code\u003e\n\u003c\/li\u003e\n  \u003cli\u003e\"Integrate legacy PHP codebase với new Node.js microservice\" → \u003ccode\u003eultrathink\u003c\/code\u003e\n\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch2\u003eNhóm 2: Pipe Output Pattern\u003c\/h2\u003e\n\n\u003ch3\u003ePattern đáng tin cậy nhất cho debugging\u003c\/h3\u003e\n\u003cp\u003eMột trong những patterns phổ biến và hiệu quả nhất từ cộng đồng:\u003c\/p\u003e\n\n\u003cpre\u003e\u003ccode\u003e# Gửi error log trực tiếp để Claude phân tích\ncat error.log | claude \"explain this error and suggest a fix\"\n\n# Fix failing tests tự động\nnpm test 2\u0026gt;\u0026amp;1 | claude \"fix the failing tests\"\n\n# CI output → fix ngay\ncat ci-output.txt | claude \"Go fix the failing CI tests\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003cp\u003eLý do pattern này mạnh: bạn gửi \u003cstrong\u003eraw data\u003c\/strong\u003e thay vì mô tả vấn đề bằng lời. Claude thấy chính xác error message, stack trace, và context — không phải version bạn \"dịch lại\" có thể thiếu thông tin quan trọng.\u003c\/p\u003e\n\n\u003ch3\u003eVariations thực tế\u003c\/h3\u003e\n\u003cul\u003e\n  \u003cli\u003ePaste URL của Pull Request: \"Go fix the failing CI tests in this PR\"\u003c\/li\u003e\n  \u003cli\u003ePaste Sentry error link: Claude đọc full stack trace và context\u003c\/li\u003e\n  \u003cli\u003ePipe database query results: \"This query is slow, optimize it\"\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch2\u003eNhóm 3: ! prefix — Command Inline trong Context\u003c\/h2\u003e\n\n\u003ch3\u003eChạy command và inject vào context ngay lập tức\u003c\/h3\u003e\n\u003cp\u003eDùng \u003ccode\u003e!\u003c\/code\u003e trước command để chạy shell command và inject output vào conversation context:\u003c\/p\u003e\n\n\u003cpre\u003e\u003ccode\u003e!git status\n!npm test\n!docker ps\n!cat package.json\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003cp\u003eCommand chạy, output land vào context, và Claude thấy ngay kết quả. Nhanh hơn rất nhiều so với \"chạy lệnh → copy output → paste vào chat.\"\u003c\/p\u003e\n\n\u003ch3\u003eUse cases đặc biệt hữu ích\u003c\/h3\u003e\n\u003cul\u003e\n  \u003cli\u003e\n\u003ccode\u003e!git log --oneline -10\u003c\/code\u003e → Claude thấy recent commits trước khi làm việc\u003c\/li\u003e\n  \u003cli\u003e\n\u003ccode\u003e!npm run build 2\u0026gt;\u0026amp;1\u003c\/code\u003e → Build errors vào context ngay\u003c\/li\u003e\n  \u003cli\u003e\n\u003ccode\u003e!psql -c \"SELECT * FROM users LIMIT 5\"\u003c\/code\u003e → Data sample cho context\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch2\u003eNhóm 4: Claude Tự Học CLI Tools Mới\u003c\/h2\u003e\n\n\u003ch3\u003ePattern: --help first, then use\u003c\/h3\u003e\n\u003cp\u003eMột trong những tricks thú vị nhất: Claude có thể tự học bất kỳ CLI tool nào, kể cả internal tools chưa từng thấy:\u003c\/p\u003e\n\n\u003cpre\u003e\u003ccode\u003e\"Use 'sentry-cli --help' to learn about it, then use it to find the most recent error in production.\"\n\n\"Use 'our-internal-tool --help' to understand its commands, then run a report for last week.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003cp\u003eClaude đọc help output, hiểu syntax, và tự chạy commands đúng. Không cần bạn biết syntax trước hay viết script wrapper.\u003c\/p\u003e\n\n\u003ch2\u003eNhóm 5: \/btw Overlay — Hỏi Nhanh Không Phá Conversation\u003c\/h2\u003e\n\n\u003ch3\u003eClarifications mid-session không interrupt flow\u003c\/h3\u003e\n\u003cp\u003e\u003ccode\u003e\/btw\u003c\/code\u003e mở pop-up overlay cho phép bạn hỏi nhanh mà không ảnh hưởng đến conversation history:\u003c\/p\u003e\n\n\u003cul\u003e\n  \u003cli\u003e\"Why did you choose this approach?\"\u003c\/li\u003e\n  \u003cli\u003e\"What's the tradeoff here?\"\u003c\/li\u003e\n  \u003cli\u003e\"Is there a simpler way to do this?\"\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003cp\u003eOverlay đóng lại, conversation tiếp tục. Không có context pollution, không mất thread. Đây là feature ít được document nhưng cực kỳ useful khi bạn cần clarity mà không muốn \"interrupt\" Claude đang trong middle of a task.\u003c\/p\u003e\n\n\u003ch2\u003eNhóm 6: CLAUDE.md — Nền Tảng Của Mọi Session\u003c\/h2\u003e\n\n\u003ch3\u003eCấu trúc CLAUDE.md hiệu quả\u003c\/h3\u003e\n\u003cp\u003eCLAUDE.md được đọc tự động khi Claude Code khởi động trong directory. Đây là nơi đặt:\u003c\/p\u003e\n\n\u003cul\u003e\n  \u003cli\u003e\n\u003cstrong\u003eProject context:\u003c\/strong\u003e Mô tả ngắn về project, stack, và architecture\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eCoding conventions:\u003c\/strong\u003e Naming conventions, formatting rules, style guide\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eForbidden actions:\u003c\/strong\u003e \"Never delete files,\" \"Always use TypeScript strict mode\"\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003ePreferred tools:\u003c\/strong\u003e \"Use axios not fetch,\" \"use pnpm not npm\"\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eEnvironment notes:\u003c\/strong\u003e Special setup, environment variables, local quirks\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch3\u003eGlobal vs Project CLAUDE.md\u003c\/h3\u003e\n\u003cul\u003e\n  \u003cli\u003e\n\u003cstrong\u003eGlobal (~\/.claude\/CLAUDE.md):\u003c\/strong\u003e Preferences áp dụng cho mọi project (coding style, general rules)\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eProject (.\/CLAUDE.md):\u003c\/strong\u003e Context cụ thể cho từng project (architecture, business rules)\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch2\u003eNhóm 7: Plan Mode — Think Before You Act\u003c\/h2\u003e\n\n\u003ch3\u003eKhi nào dùng Plan Mode\u003c\/h3\u003e\n\u003cp\u003eVới tasks phức tạp hoặc có rủi ro cao, luôn dùng Plan Mode trước:\u003c\/p\u003e\n\n\u003cpre\u003e\u003ccode\u003e# Yêu cầu Claude plan trước khi thực thi\n\"Plan how you'll refactor this authentication system, then wait for my approval before making any changes.\"\n\n# Hoặc prefix với \"think hard about\"\n\"Think hard about the best approach to migrate this database schema, then show me the plan.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003cp\u003eBenefits của Plan Mode:\u003c\/p\u003e\n\u003cul\u003e\n  \u003cli\u003ePhát hiện potential issues trước khi code được viết\u003c\/li\u003e\n  \u003cli\u003eCơ hội redirect nếu approach sai\u003c\/li\u003e\n  \u003cli\u003eBetter understanding của scope thực tế\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch2\u003eNhóm 8: Subagents và Parallel Execution\u003c\/h2\u003e\n\n\u003ch3\u003eChia task cho multiple agents\u003c\/h3\u003e\n\u003cp\u003eVới tasks lớn hoặc có thể parallelize, Claude Code hỗ trợ subagent delegation:\u003c\/p\u003e\n\n\u003cpre\u003e\u003ccode\u003e\"You need to update 5 components to use the new design system. Start 5 subagents working in parallel, each handling one component.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003cp\u003ePattern này đặc biệt hiệu quả với:\u003c\/p\u003e\n\u003cul\u003e\n  \u003cli\u003eBatch migrations (update nhiều files với cùng pattern)\u003c\/li\u003e\n  \u003cli\u003eMulti-file refactoring\u003c\/li\u003e\n  \u003cli\u003eTest generation cho nhiều modules\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003cp\u003eXem thêm tại \u003ca href=\"\/products\/chief-of-staff-agent-di%E1%BB%81u-ph%E1%BB%91i-multi-agent-v%E1%BB%9Bi-claude-sdk\"\u003eChief of Staff Agent — Điều phối multi-agent với Claude SDK\u003c\/a\u003e.\u003c\/p\u003e\n\n\u003ch2\u003eNhóm 9: Worktrees cho Parallel Development\u003c\/h2\u003e\n\n\u003ch3\u003eMultiple Claude instances, không conflict\u003c\/h3\u003e\n\u003cp\u003eGit worktrees + Claude Code = parallel development không cần branch switching:\u003c\/p\u003e\n\n\u003cpre\u003e\u003ccode\u003e# Tạo worktree cho feature mới\ngit worktree add ..\/project-feature-auth feature\/auth\n\n# Mở Claude Code trong worktree riêng\ncd ..\/project-feature-auth \u0026amp;\u0026amp; claude\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003cp\u003eMỗi worktree có Claude instance riêng, working độc lập trên branch riêng. Không conflict, không phải stash\/unstash liên tục.\u003c\/p\u003e\n\n\u003ch2\u003eNhóm 10: Keyboard Shortcuts Tiết Kiệm Thời Gian\u003c\/h2\u003e\n\n\u003cp\u003eMột số shortcuts quan trọng:\u003c\/p\u003e\n\u003cul\u003e\n  \u003cli\u003e\n\u003ckbd\u003eCtrl+C\u003c\/kbd\u003e — Interrupt current generation (không exit, chỉ stop)\u003c\/li\u003e\n  \u003cli\u003e\n\u003ckbd\u003eCtrl+L\u003c\/kbd\u003e — Clear screen, giữ context\u003c\/li\u003e\n  \u003cli\u003e\n\u003ckbd\u003e↑\u003c\/kbd\u003e — Previous command trong history\u003c\/li\u003e\n  \u003cli\u003e\n\u003ckbd\u003eTab\u003c\/kbd\u003e — Autocomplete file paths và commands\u003c\/li\u003e\n  \u003cli\u003e\n\u003ccode\u003e\/compact\u003c\/code\u003e — Compress conversation history khi dài\u003c\/li\u003e\n  \u003cli\u003e\n\u003ccode\u003e\/clear\u003c\/code\u003e — Reset conversation hoàn toàn\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch2\u003eNhóm 11: Hooks — Tự Động Hóa Tasks Lặp Lại\u003c\/h2\u003e\n\u003cp\u003eClaude Code hooks cho phép chạy scripts tự động tại specific points trong workflow:\u003c\/p\u003e\n\n\u003cpre\u003e\u003ccode\u003e# Ví dụ hook configuration trong .claude\/hooks\/\n{\n  \"pre_tool_use\": [\".\/scripts\/check-permissions.sh\"],\n  \"post_tool_use\": [\".\/scripts\/log-changes.sh\"],\n  \"on_session_start\": [\".\/scripts\/load-context.sh\"]\n}\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003cp\u003eUse cases phổ biến cho hooks:\u003c\/p\u003e\n\u003cul\u003e\n  \u003cli\u003e\n\u003cstrong\u003ePre-commit validation:\u003c\/strong\u003e Tự động chạy linter trước mỗi commit Claude thực hiện\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eAuto-backup:\u003c\/strong\u003e Backup files trước khi Claude modify\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eContext loading:\u003c\/strong\u003e Tự động inject project context vào mỗi session\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eNotification:\u003c\/strong\u003e Alert khi Claude hoàn thành long-running task\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eSecurity checks:\u003c\/strong\u003e Verify không có sensitive data trong files sắp được committed\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch2\u003eNhóm 12: Model Selection — Chọn Model Phù Hợp\u003c\/h2\u003e\n\u003cp\u003eClaude Code support nhiều models với tradeoffs khác nhau:\u003c\/p\u003e\n\n\u003cul\u003e\n  \u003cli\u003e\n\u003cstrong\u003eClaude Opus (most capable):\u003c\/strong\u003e Tốt nhất cho complex reasoning, architecture decisions, legacy code. Chi phí cao nhất, chậm nhất.\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eClaude Sonnet (balanced):\u003c\/strong\u003e Sweet spot cho hầu hết coding tasks. Good quality, reasonable speed và cost.\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eClaude Haiku (fastest\/cheapest):\u003c\/strong\u003e Cho simple tasks: rename files, add comments, format code. Không waste expensive model cho simple tasks.\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003cp\u003e\u003cstrong\u003ePractical strategy:\u003c\/strong\u003e Default Sonnet, upgrade Opus khi cần deep thinking (dùng \u003ccode\u003eultrathink\u003c\/code\u003e), downgrade Haiku cho batch simple operations.\u003c\/p\u003e\n\n\u003ch2\u003eNhóm 13: Git Integration Best Practices\u003c\/h2\u003e\n\u003cp\u003eClaude Code có native git understanding. Tips để leverage tốt nhất:\u003c\/p\u003e\n\n\u003ch3\u003eCommit granularity\u003c\/h3\u003e\n\u003cpre\u003e\u003ccode\u003e# Tốt: Yêu cầu Claude commit theo logic units\n\"Implement the login flow. Commit each component separately:\n1. Commit the auth middleware\n2. Commit the login route handler\n3. Commit the frontend form component\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch3\u003eBranch strategy\u003c\/h3\u003e\n\u003cpre\u003e\u003ccode\u003e# Luôn làm việc trên branch, không trực tiếp trên main\n\"Create a new branch 'feature\/user-auth' and implement authentication there.\nDo NOT push to main until all tests pass.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch3\u003eRollback safety\u003c\/h3\u003e\n\u003cpre\u003e\u003ccode\u003e# Yêu cầu Claude stash trước khi làm gì risky\n\"Before refactoring this module, create a git stash named 'pre-refactor-backup'\nso we can rollback if needed.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch2\u003eNhóm 14: Context Management Advanced\u003c\/h2\u003e\n\n\u003ch3\u003eKhi nào dùng \/compact vs \/clear\u003c\/h3\u003e\n\u003cul\u003e\n  \u003cli\u003e\n\u003cstrong\u003e\/compact:\u003c\/strong\u003e Giữ key context nhưng compress. Dùng khi conversation dài nhưng vẫn cần context từ trước.\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003e\/clear:\u003c\/strong\u003e Reset hoàn toàn. Dùng khi chuyển sang completely different task trong cùng session.\u003c\/li\u003e\n\u003c\/ul\u003e\n\n\u003ch3\u003eStrategic context injection\u003c\/h3\u003e\n\u003cp\u003eVới long projects, start mỗi session bằng context primer:\u003c\/p\u003e\n\u003cpre\u003e\u003ccode\u003e\"Project context:\n- Stack: Next.js 14, TypeScript, PostgreSQL, Prisma\n- Current goal: implement user dashboard\n- Completed: auth, user model, basic CRUD\n- In progress: dashboard UI\n- Next: analytics components\nLet's continue with the analytics components.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch3\u003eFile references vs paste\u003c\/h3\u003e\n\u003cp\u003eKhi cần Claude đọc file, reference file path thay vì paste content khi có thể — tiết kiệm context space:\u003c\/p\u003e\n\u003cpre\u003e\u003ccode\u003e\"Read src\/utils\/auth.ts và explain the token validation logic\"\n# Tốt hơn paste toàn bộ file vào chat\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch2\u003eNhóm 15: Debugging Strategies\u003c\/h2\u003e\n\n\u003ch3\u003eBinary search debugging\u003c\/h3\u003e\n\u003cpre\u003e\u003ccode\u003e\"This test suite has 30 failing tests. Identify which 15 could be related,\nthen narrow down to the root cause systematically.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch3\u003eRubber duck debugging với Claude\u003c\/h3\u003e\n\u003cpre\u003e\u003ccode\u003e\"I'll explain my understanding of why this code should work:\n[explanation]\nPoint out the flaw in my reasoning.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch3\u003eComparative debugging\u003c\/h3\u003e\n\u003cpre\u003e\u003ccode\u003e\"Here's the code that works: [code A]\nHere's the code that doesn't: [code B]\nThey should do the same thing. Find the difference causing the bug.\"\u003c\/code\u003e\u003c\/pre\u003e\n\n\u003ch2\u003eBonus: 5 Mistakes Phổ Biến Nhất\u003c\/h2\u003e\n\u003col\u003e\n  \u003cli\u003e\n\u003cstrong\u003eKhông có CLAUDE.md:\u003c\/strong\u003e Claude bắt đầu mỗi session \"mù\" về project context\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003ePrompt quá vague:\u003c\/strong\u003e \"Fix the bug\" vs \"Fix the TypeError in user-auth.js line 47 where req.user is undefined\"\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eUltrathink mọi thứ:\u003c\/strong\u003e Tốn token không cần thiết cho simple tasks\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eKhông review AI-generated code:\u003c\/strong\u003e Claude code thường đúng nhưng không phải luôn luôn optimal\u003c\/li\u003e\n  \u003cli\u003e\n\u003cstrong\u003eKhông commit sau mỗi working state:\u003c\/strong\u003e Claude có thể \"go down a rabbit hole\" — checkpoint thường xuyên\u003c\/li\u003e\n\u003c\/ol\u003e\n\n\u003cp\u003eXem thêm hướng dẫn toàn diện tại \u003ca href=\"\/products\/claude-code-toan-t%E1%BA%ADp-l%E1%BA%ADp-trinh-v%E1%BB%9Bi-ai-agent-trong-terminal\"\u003eClaude Code toàn tập — Lập trình với AI agent trong terminal\u003c\/a\u003e và \u003ca href=\"\/products\/claude-md-masterclass-t%E1%BB%91i-%C6%B0u-workspace-cho-claude-code\"\u003eCLAUDE.md Masterclass\u003c\/a\u003e.\u003c\/p\u003e\n\n\u003ch2\u003eNguồn tham khảo\u003c\/h2\u003e\n\u003cul\u003e\n  \u003cli\u003e\u003ca href=\"https:\/\/www.builder.io\/blog\/claude-code-tips-best-practices\" target=\"_blank\" rel=\"noopener\"\u003eBuilder.io — 50 Claude Code Tips and Best Practices (Vishwas Gopinath, Mar 2026)\u003c\/a\u003e\u003c\/li\u003e\n\u003c\/ul\u003e\n","brand":"Minh Tuấn","offers":[{"title":"Default Title","offer_id":47725832634580,"sku":null,"price":0.0,"currency_code":"VND","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/0821\/0264\/9044\/files\/50-m_o-claude-code-th_c-chi_n-t_-thinking-tokens-d_n-subagents.jpg?v=1774579743","url":"https:\/\/claude.vn\/products\/50-m%e1%ba%b9o-claude-code-th%e1%bb%b1c-chi%e1%ba%bfn-t%e1%bb%ab-thinking-tokens-d%e1%ba%bfn-subagents","provider":"CLAUDE.VN","version":"1.0","type":"link"}